模型卡、文件列表和校验和已在此收录。如果你持有文件并有权分享,可以提交种子,让其他人从节点下载。
获取来源文件需要 Hugging Face 批准。The abliterated (refusal-removed) build of Nex-AGI's Nex-N2.5-mini — baked directly into the official BF16 shards, byte-for-byte drop-in for the original
|
🛡️ New from Orca — OrcaCyber Zero 1.5 Post-trained for cybersecurity: vulnerability analysis, vulnerability reproduction, red teaming & offensive-security research. 98.01% pass@1 on CyberGym Level 1 · 1M context · function calling · OpenAI-compatible API. |
|
|
🔌 New from Orca — Orca Connect Have a rig running Ollama, LM Studio, vLLM, llama.cpp or SGLang? Orca Connect puts it behind your usual OrcaRouter API key, next to 200+ hosted models, reachable from a laptop, a phone or a cloud agent. Outbound only: no port forwarding, no tunnel, no firewall rule. Requests your own hardware serves cost $0, and |
|
One Gateway. Every Model. — Route Smarter · Ship Safer · Spend Less.
Website · Model Catalog · GitHub · Discord · X
The abliterated (refusal-removed) build of
nex-agi/Nex-N2.5-mini— a ~34B Mixture-of-Experts agentic VLM with hybrid gated-delta-net + full attention, 256 routed experts top-8 beside a shared expert on every layer, a native vision tower, and a 262K-token context.The refusal direction is baked directly into the official BF16 shards — same dtype, same shard layout, same
model.safetensors.index.json. All 1026 tensors match the base checkpoint in name, dtype and shape, so this is a drop-in replacement fornex-agi/Nex-N2.5-miniin any stack that already serves it.On precision: BF16 is the format Nex-AGI ships Nex-N2.5-mini in — there is no quantization step here, applied or undone.
This model has had its safety alignment substantially removed via abliteration (orthogonalizing the refusal direction out of the residual stream). As a direct consequence:
Nex-N2.5-mini would refuse. It has no meaningful built-in guardrails.By downloading or using this model you acknowledge and accept the above.
Models are only half the system.
OrcaCode Review turns every model listed on OrcaRouter into a production code-review agent:
Open model. Open harness. Open bill.
| Base model | nex-agi/Nex-N2.5-mini |
| Architecture | Qwen3_5MoeForConditionalGeneration (qwen3_5_moe) — 40 layers, hidden 2048, hybrid attention (30 gated-delta-net linear-attention + 10 full-attention with attn_output_gate, interval 4), 256 routed experts top-8 + 1 shared expert on every layer, interleaved mRoPE, native vision tower (27 blocks @ 1152, patch 16, spatial merge 2) |
| Parameters | ~34B total — 35.1B tensor elements on disk, including the 0.44B vision tower |
| Modification | Abliteration (refusal-direction removal), baked on disk into the BF16 shards — no quantization applied |
| Format | safetensors, BF16, 16 shards, 1026 tensors, 66 GiB |
| Edited | 122 of 1026 tensors; the other 904 are byte-identical to the base |
| Preserved | Everything that reads or gates the residual stream rather than writing it — the MoE router, the expert and attention input projections, the linear-attention state parameters, every norm, lm_head, and the vision tower apart from the merger's output projection |
| Context | 262,144 tokens |
| Vocabulary | 248,320 (untied embeddings) |
| Recommended for | Red-team & refusal-mechanism research, robustness evaluation, and as a base for further post-training / quantization |
Refusal-direction removal following Arditi et al. (2024), Refusal in Language Models Is Mediated by a Single Direction. A single refusal direction is estimated from the residual stream on paired harmful/harmless prompts, then projected out of the matrices that write that stream — the attention output projections, the routed and shared expert down-projections, the token embeddings, and the vision merger's output projection. The projection is computed in float32 and stored back in the checkpoint's own BF16.
122 of the 1026 tensors are edited; the other 904 are byte-identical to the base. Nothing that reads or gates the residual stream is touched. That is verified by a tensor-by-tensor comparison of the two checkpoints rather than by trusting the code that produced them, which is also what guarantees the drop-in property: every dtype, shape and shard boundary survives.
The edit is applied to the whole depth of the model, and the layer the direction is fitted at was selected on a held-out validation split by an objective that rewards bypassing refusal while charging for both divergence on harmless prompts and any newly induced refusal of benign ones. Capability retention is therefore measured rather than assumed — see Evaluation.
Nothing here changes the architecture, the config, the tokenizer or the shard layout, so anything that fine-tunes or quantizes the base checkpoint works unmodified on this one. Note that further training on refusal-bearing data will partially restore the direction; re-derive and re-bake afterwards rather than assuming the edit survives.
Measured on these exact weights, 8×H100, greedy decoding, 64 new tokens, reasoning_effort=none,
thinking off. Both columns come from the same script, the same sampling seed, the same settings and the
same scoring code. Refusal is judged by a rule-based opening-phrase classifier (caveat = answered but
wrapped in a disclaimer) — indicative, not an LLM-judge / publication-grade number. Read
Limitations before quoting the 0.000s.
| Benchmark | n | Base | This model | Δ | Caveat (base → this) |
|---|---|---|---|---|---|
| AdvBench | 100 | 1.000 | 0.000 | −100% | 1.000 → 0.360 |
| JailbreakBench (harmful) | 100 | 0.980 | 0.000 | −98% | 0.980 → 0.370 |
| StrongREJECT | 150 | 0.933 | 0.000 | −93% | 0.953 → 0.360 |
| MaliciousInstruct | 100 | 0.950 | 0.000 | −95% | 0.950 → 0.400 |
| HarmBench (standard) | 150 | 0.980 | 0.020 | −96% | 0.980 → 0.227 |
| ForbiddenQuestions | 150 | 0.587 | 0.020 | −57% | 0.627 → 0.273 |
| SimpleSafetyTests | 50 | 0.960 | 0.060 | −90% | 0.960 → 0.380 |
The base model refuses 93–100% of harmful prompts on six of the seven sets. ForbiddenQuestions is the
outlier at 0.587, which says more about that set — a large share of its questions are sensitive rather
than harmful, and the base model answers them — than about the edit.
The caveat column is the second half of the story. It does not merely track refusal: it falls from ~0.96 to ~0.35, so roughly two thirds of the disclaimer-wrapping goes with the refusal rather than surviving it.
| Benchmark | n | Base | This model | Δ |
|---|---|---|---|---|
| XSTest-safe | 250 | 0.056 | 0.008 | −4.8 pp |
| JailbreakBench (benign) | 100 | 0.220 | 0.000 | −22.0 pp |
XSTest-safe is 250 prompts written to look harmful while being benign. The base model is already fairly restrained there at 5.6%; the edit does not blunt that discrimination, it sharpens it. JBB's benign split is the harder one — adversarially benign, 22% wrongly refused at baseline — and it goes to zero.
| Benchmark | n | Base | This model | Δ |
|---|---|---|---|---|
| MMLU (0-shot) | 300 | 0.830 | 0.830 | ±0.0 pp |
| MMLU-Pro (CoT) | 250 | 0.776 | 0.752 | −2.4 pp |
| GSM8K (CoT) | 150 | 0.907 | 0.907 | ±0.0 pp |
| CMMLU (0-shot, Chinese) | 500 | 0.852 | 0.864 | +1.2 pp |
Every delta is within ±2.4 pp, and unparsed rates were ≤1.6% on both checkpoints. Two of the four move up or hold exactly, which is the signature of sampling noise at these sizes rather than of an improvement — read the table as "unchanged", not as a gain. The −2.4 pp on MMLU-Pro is the largest single move and sits within the noise band of a 250-item CoT subset; it is not evidence of a regression, and it is not evidence against one either.
MMLU and CMMLU ask for a bare option letter, so they are scored from a single forward pass — an argmax over the option letters' logits. MMLU-Pro and GSM8K are chain-of-thought scored by generation, because their prompts ask the model to reason first: an argmax at the first position there measures the opening of a chain of thought, not an answer, and reads ~55 points low. These MMLU-Pro figures are therefore directly comparable to CoT-scored MMLU-Pro results elsewhere.
qwen3_5_moe needs transformers ≥ 5.2; verified here on 5.17.0 with torch 2.14. Install
flash-linear-attention as well — without it
transformers falls back to a reference PyTorch gated-delta-rule on 30 of the 40 layers, which is correct
but much slower.
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "orcarouter/Nex-N2.5-mini-Uncensored"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="bfloat16", device_map="auto")
msgs = [{"role": "user", "content": "..."}]
text = tok.apply_chat_template(
msgs, tokenize=False, add_generation_prompt=True, reasoning_effort="none"
)
out = model.generate(**tok(text, return_tensors="pt").to(model.device), max_new_tokens=512)
print(tok.decode(out[0], skip_special_tokens=True))
The weights are 66 GiB, so they fit on a single 80 GB card with room for activations, or shard across
smaller ones with device_map="auto".
Serving: we did not serve this build through vLLM or SGLang, so no flag set is recommended here. Because every tensor matches the base checkpoint in name, dtype and shape, whatever recipe works for
nex-agi/Nex-N2.5-miniapplies unchanged — follow the base model's own serving instructions.
Nex-N2.5-mini has no enable_thinking toggle — passing one is silently inert, because
apply_chat_template drops unknown kwargs into a non-strict jinja context and the prompt renders
byte-identically either way. Control it with reasoning_effort:
| value | renders | use |
|---|---|---|
none |
<think>\n\n</think>\n\n — pre-closed |
the reply starts at the answer; this family's thinking-off, and what every number above was measured with |
| unset / other | <think> |
reasoning on, template default |
high |
<think>\n |
reasoning on, longer traces |
none is not a quieter low: it closes the thinking block inside the prompt, so no reasoning budget is
needed and the reply is the answer. With reasoning on, give generation enough budget to reach </think>
or replies get truncated inside the scratchpad.
Note for evaluators: do not pair
reasoning_effort="none"with a<think>-stripping classifier. The block is already closed in the prompt, so the reply contains no</think>, every reply strips to the empty string, and the run reports a 100% refusal rate on every split.
Nex-N2.5-mini.visual.merger.linear_fc2 writes the language residual
stream, so an image could otherwise re-introduce the direction the text writers no longer carry, which
is why it is included. But the direction was fitted on text only, and on transformers 5.17
AutoModelForImageTextToText does not register this config — the checkpoint loads text-only as
Qwen3_5MoeForCausalLM — so no image prompt could be pushed through to check the effect. Treat
image-conditioned behaviour as untested in both directions: neither confirmed removed nor confirmed
intact.Apache 2.0, inherited from the base model
nex-agi/Nex-N2.5-mini. Abliteration does not change the
underlying license obligations.
model-00006-of-00016.safetensors 5.0 GBmodel-00009-of-00016.safetensors 5.0 GBmodel-00012-of-00016.safetensors 5.0 GBmodel-00015-of-00016.safetensors 5.0 GBmodel-00003-of-00016.safetensors 5.0 GBmodel-00001-of-00016.safetensors 4.7 GBmodel-00007-of-00016.safetensors 4.5 GBmodel-00010-of-00016.safetensors 4.5 GBmodel-00013-of-00016.safetensors 4.5 GBmodel-00004-of-00016.safetensors 4.5 GBmodel-00008-of-00016.safetensors 4.0 GBmodel-00011-of-00016.safetensors 4.0 GBmodel-00014-of-00016.safetensors 4.0 GBmodel-00002-of-00016.safetensors 4.0 GBmodel-00005-of-00016.safetensors 4.0 GBmodel-00016-of-00016.safetensors 2.7 GB